自主定位是移动机器人领域的关键核心技术。目前煤矿井下机器人自主定位能力尚显不足。近年来，随着智能图像处理技术的快速发展，视觉定位技术展现出广阔的改进空间和强大的适应性，在自动驾驶等领域得到了广泛应用。在此背景下，煤矿井下视觉定位技术也受到日益广泛的关注。论文针对煤矿井下视觉定位中的技术难题，提出了适用于井下移动机器人定位的视觉感知系统架构，并深入研究了深度图像误差补偿和空洞修复、煤矿井下低光图像增强、纹理退化场景图像配准和位姿估计、以及动态场景运动分割与静态背景重建方法，旨在改善因深度传感器误差、光照不足或光照不均、纹理退化、以及场景中运动目标引起的性能下降，为定位系统提供准确和稳定的视觉信息。主要研究工作如下：

⑴ 在深入分析煤矿井下场景特点的基础上，提出了面向井下移动机器人定位的视觉感知系统架构。该架构采用RGB-D相机作为前端传感器，主要包括深度图像预处理、彩色图像预处理、动态预处理、视觉里程计等模块。其中，深度图像预处理对RGB-D 相机获取的深度图像进行误差补偿和空洞修复，为视觉定位提供准确的空间距离信息；彩色图像预处理对RGB-D 相机获取的彩色图像进行增强，为视觉里程计提供光照均匀纹理清晰的图像；动态预处理识别并滤除场景中的动态特征并重建静态背景，为定位和建图提供完整的静态场景信息；视觉里程计完成图像特征提取与匹配，并实时估计相机位姿，实现视觉定位。

⑵ 针对现有深度图像误差补偿方法难以兼顾速度和精度的问题，提出一种基于拟合优化的深度误差补偿方法，首先建立三维误差查找表并对其进行曲线拟合，构建误差模型；然后利用TOF深度误差的旋转对称性和分布规律优化误差模型。所提出的优化策略大幅减少了模型参数，在保持补偿精度的同时提升了效率。针对现有深度图像空洞修复方法在处理图像中的跨边界空洞和大面积空洞时出现的边界失真和高迭代次数问题，提出一种基于边缘重建的自适应曲率扩散修复方法。首先利用彩色图像与深度图像之间的共生性重建缺失的深度边缘，有效解决了跨边界空洞修复中的边界失真问题；然后，在曲率驱动扩散模型中引入梯度引导函数，根据边缘梯度的模值自适应调整扩散强度，实现了对大面积内部空洞的快速修复。在煤矿井下RGB-D数据集上验证了所提出方法修复井下深度图像的有效性。

⑶ 针对现有图像增强方法在处理煤矿井下光照不均匀的低光图像时易出现局部过增强、光晕和伪影的问题，提出一种基于局部感知的低光图像自适应增强方法。设计了多尺度有效引导滤波器和自适应亮度校正函数，并将它们分别应用于光照估计、亮度校正和反射增强过程中，实现了对光照、纹理和噪声的局部感知和自适应增强。在煤矿井下图像数据集上，通过图像增强实验、图像增强前/后特征提取和匹配对比实验，验证了所提出方法在视觉定位任务中的有效性。

⑷ 针对特征点法视觉里程计在煤矿井下纹理退化场景中因难以有效提取和匹配特征点而导致定位性能下降的问题，提出一种基于点线特征一致性约束的视觉里程计。首先，对ORB（Orien-ted FAST And Rotated BRIEF）算法的角点提取策略进行优化，通过引入动态阈值和多尺动态四叉树等方法，提升点特征在纹理退化和光照变化区域的帧间稳定性和分布均匀性。然后，对GMS（Grid-based Motion Statistics）算法进行改进，通过引入动态网格划分和高斯权重因子增强了全局运动一致性约束，提升了纹理退化区域中特征匹配的准确性。最后，在位姿估计中引入EDlines线特征，增强对纹理退化场景的结构感知能力。在TUM公共数据集和煤矿井下图像序列中进行了点/线特征匹配实验、视觉定位实验和煤矿井下场景实验，验证了所提出方法在纹理退化场景中的有效性和鲁棒性。

⑸ 针对煤矿井下动态场景中因动态目标干扰而导致定位和建图性能下降的问题，首先改进现有的联合语义-几何约束的运动分割方法，通过形态学方法优化语义分割边缘，并将运动一致性原理和多视图几何分析相结合来检测目标运动状态，实现了对具有不同运动特性的多样化运动目标的准确识别和分割；然后，分别利用光流约束和几何约束对彩色图像和深度图像进行静态背景重建，改善了稠密建图中的动态空洞或动态重影问题。此外，针对井下移动机器人计算资源受限的问题，设计了一种基于边缘-终端计算的视觉同步定位与建图部署架构，通过终端与边缘服务器的分布式并行处理，提升定位及建图的执行效率。在TUM公共数据集和煤矿井下图像序列上，进行了动态场景定位实验、静态背景重建和稠密建图实验、煤矿井下动态场景定位实验以及终端-边缘架构仿真实验，证明了所提出方法在煤矿井下动态场景中的有效性。

本文提出的煤矿井下视觉感知架构及系列改进方法能够有效增强机器人对低光照、纹理退化和动态工作环境的感知能力，显著提升定位的准确性和鲁棒性。研究成果为煤矿井下视觉定位技术的应用提供了理论和技术支撑，对于推动煤矿装备智能化和机器人化转型的研究与应用具有积极意义。

Autonomous localization stands as a pivotal technology within the domain of mobile robotics. Despite considerable progress, the localization capabilities of robots in underground coal mines remain suboptimal, substantially hindering the roboticization of mobile equipment in these settings. The recent rapid advancement of intelligent image processing technologies has significantly enhanced the potential for improvement and robust adaptability of visual positioning technology, which has been widely applied in fields such as autonomous driving. Consequently, the application of visual positioning technology in underground coal mines has has attracted increasing attention. In this dissertation, a visual perception system architecture for the localization of mobile robots has proposed. The research includes depth image error compensation and hole filling, low-light image enhancement specific to underground coal mines, image registration and pose estimation in scenes with texture degradation, and methods for dynamic scene motion segmentation and static background reconstruction. The primary objective is to enhance the accuracy and stability of visual information in low-light, texture-degraded, and dynamic scenes, thereby improving the performance of visual localization systems. The main research content is as follows:

⑴ Based on an exhaustive analysis of the characteristics of the underground coal mine environment, a visual perception system architecture tailored for mobile robot localization is been proposed. This architecture, with an RGB-D camera as the primary sensor, includes key modules such as depth image preprocessing, color image preprocessing, dynamic scene preprocessing, and visual odometry. The depth image preprocessing module corrects errors and fills holes in the depth images from the RGB-D camera, providing accurate distance information crucial for precise visual positioning. The color image preprocessing module enhances the RGB-D camera's color images, ensuring that the images fed into the visual odometry are well-lit with distinct textures. The dynamic scene preprocessing module identifies and segments moving objects, eliminating their dynamic features. Furthermore, it replaces the dynamic regions in the image frames with a static background, thereby providing a complete static scene for localization and mapping. The visual odometry module performs feature extraction and matching from images and estimates the camera pose in real-time, achieving high-precision visual positioning.

⑵ To address the challenge of balancing speed and accuracy in existing depth image error compensation methods, a fitting optimization-based depth error compensation method is proposed. This approach begins by constructing a three-dimensional error lookup table and employs curve fitting to establish an error model. It subsequently refines the model by exploiting the rotational symmetry and distribution patterns of edge distortion errors. The proposed optimization strategy markedly reduces model parameters, thereby enhancing efficiency without compromising the accuracy of the compensation. Furthermore, to tackle the boundary distortion and high iteration counts in existing depth image hole filling methods, particularly for cross-boundary and large-area holes, we introduce an edge reconstruction-based adaptive curvature diffusion method. This method reconstructs missing depth edges by leveraging the correlation between color and depth images, effectively resolving boundary distortion in cross-boundary hole filling. It then integrates a gradient guidance function into the curvature-driven diffusion model, adjusting the diffusion intensity based on edge gradient magnitudes to facilitate the swift repair of large internal holes. The efficacy of these methods has been confirmed through validation on an RGB-D dataset from underground coal mines.

⑶ To address the challenges of local over-enhancement, halo effects, and artifacts when enhancing low-light images with uneven illumination typically found in underground settings, an adaptive low-light image enhancement method based on local perception is proposed. This method incorporates a multi-scale effective guidance filter designed to accurately perceive local illumination and texture within images. Additionally, an adaptive brightness correction function has been developed to boost brightness in low-light areas while curbing over-enhancement in high-light regions. The efficacy of the proposed method in visual localization tasks has been established through image enhancement experiments and by comparing feature extraction and matching results before and after enhancement, utilizing an underground coal mine image dataset.

⑷ To counter the diminished positioning accuracy of feature-based visual odometry in texture-degraded environments, a visual odometry approach based on point-line feature consistency constraints is proposed. This approach begins by optimizing the corner extraction strategy of the ORB (Oriented FAST And Rotated BRIEF) algorithm through the implementation of dynamic thresholds and a multi-scale dynamic quadtree. This optimization enhances the inter-frame stability and uniform distribution of point features, particularly in regions affected by texture degradation and fluctuating lighting conditions. Building on this, the GMS (Grid-based Motion Statistics) algorithm is enhanced by introducing dynamic grid partitioning and Gaussian weight factors, which reinforce global motion consistency constraints and, consequently, improve the precision of feature matching in texture-degraded areas. To further augment the method's structural perception capabilities in such scenarios, EDlines line features are incorporated into the pose estimation process. The efficacy and robustness of this proposed method were substantiated through a series of experiments, including point/line feature matching, visual positioning, and underground scene analyses, conducted using both the TUM public dataset and image sequences from underground coal mines.

⑸ To address the degradation in localization and mapping performance within dynamic underground coal mine environments, firstly, a improved motion segmentation method that integrates semantic and geometric constraints is proposed. By employing morphological methods to refine the edges of semantic segmentation and combining the principle of motion consistency with multi-view geometric analysis to detect the motion states of targets, accurate identification and segmentation of moving targets with diverse motion characteristics have been achieved. Subsequently, static background reconstruction for both color and depth images was carried out using optical flow constraints and geometric constraints, respectively, which ameliorated the issues of dynamic holes or ghosting in dense mapping. Furthermore, considering the limited computational resources of underground mobile robots, an edge-terminal based visual simultaneous localization and mapping (SLAM) deployment architecture has been proposed. This architecture enhances the execution efficiency of localization and mapping through distributed parallel processing between terminals and edge servers. Experiments conducted on the TUM public dataset and underground coal mine image sequences, including dynamic scene localization, static background reconstruction, dense mapping, and edge-cloud architecture simulations, have validated the effectiveness of the proposed methods in dynamic underground scenarios.

The visual perception architecture and series of improvement methods proposed in this paper can effectively enhance the robot's perception capabilities in low-light, texture-degraded, and dynamic environments, significantly improving the accuracy and robustness of localization. Furthermore, the research findings offer both theoretical and technical underpinnings for the deployment of visual positioning technology in underground coal mining settings, holding substantial promise for advancing the intelligent and robotic transformation of coal mining equipment.